santa hat
deerdeer nosedeer glow
Sign In

Basic tips for upscale images

Basic tips for upscale images

In this article I write about my own experience on quite basic way of upscaling images.

I usually generate with ponyDiffusion model - its very versatile, it can do various poses and characters - but it gives you very bad and distorted faces if its not a super close-up. Of cause that could be fixed by upscale, but there's another problem - upscale on SDXL models can take quite some time and requires A LOT of VRAM. And in the end resuld won't be ideal.

So I saw some guys upscale originaly made with SDXL-model images with SDv1.5 models. SDv1.5 can easily upscale up to 3 times, it uses much less VRAM, and the process is quicker. Also modern SDv1.5 models are very good with adding small details to image. When you upscale image from 768*768px, on which you can hardly see eyes or other facial features on cowboy-shot image, to 2-3 times you can see sometimes light facial hair on women's face, if you use good loras.

So I've tried to upscale pony-made images from "safe" resolution 904*1160px by x1.5...x3 and got quite good results - eyes and face in general much better, lines are more pronounced and overal "sense" of and image is intact.

I guess that's a well-known way.

In this post is an example https://civitai.com/posts/2528174

12

Comments